7 Things They Won’t Tell You About Installing a Chatbot
Artificial intelligence (AI) is causing a buzz throughout the contact centre industry, in the way that IVR and outsourcing did previously.
Leading this charge is undoubtedly the chatbot. Why? The unfortunate answer is likely that they are the easiest idea to sell.
It’s easy to comprehend the benefits of what chatbots bring in reducing contact volumes, but in practice they are a very tricky technology to implement.
With this in mind, here are seven things that you need to know about chatbots before you consider installing them in your contact centre.
1. Chatbots Are Not “Plug in and Play”
In terms of ease of implementation, a chatbot is arguably the most testing technology to install in the contact centre – especially if you want a good one.
Thomas Hebner, Head of Product Innovation, Voice Technology and AI at Nuance Communications, says: “The thing I’ve been most impressed with in conversational AI is the audacity of marketers, over-hyping what ‘robots’ can do.”
“I’ve spent the last year walking into contact centres and saying: ‘sorry, you’re not going to have a bot that you can just plug in and play’.”
Thomas’s point here is that good chatbots are complex and require a lot of work. Just think about it. There’s a whole discipline around ‘conversational experience design’, with specialists in psycholinguistics backgrounds whose job it is to build a conversation around a process to teach a robot how to have a conversation. It takes a long time.
As Dr David Naylor, the Founder of Humanotics, says: “A bank in Australia came across almost 2,000 different ways for how people could ask for a bank balance. These were all programmed into their chatbot. You would think this query would be straightforward, but it takes time!”
However, it seems that organisations are finally recognising that if they want a good chatbot, they will need to invest a lot of time and data.
2. Many of the Chatbots on the Market Contain No AI
With the buzz around AI, we’ve seen an explosion of bot-building platforms. There are hundreds of new start-ups saying ‘build a bot in 30 minutes or an hour’, which sounds fantastic. But those bots end up becoming simple question-and-answer bots, essentially an expensive FAQ system.
So, if you come across a vendor that says that they can get a chatbot on your site in a matter of days, hours or even minutes, walk away! It will almost certainly not contain any “real AI” at all.
There are bots that feed a customer’s query as text into a natural language processing (NLP) system that comes back with a code that signifies an intent, so that the bot can provide a scripted response.
Then there are bots that feed a customer’s query as text into a natural language processing (NLP) system that comes back with a code that signifies an intent, so that the bot can provide a scripted response.
These can work well in certain, predefined circumstances, but the script needs to be extremely robust, requiring constant tuning, and this type of chatbot will only be able to respond to your last request, i.e. it won’t be able to “remember” what you’ve asked previously.
Creating a chatbot in this way is something Thomas Hebner does not recommend, stating: “It shouldn’t take a human to talk to another human to understand a business problem to teach a robot how to solve that problem – that’s insane!”
“The most sophisticated chatbots are being made to auto-learn these conversations in the contact centre, so that job becomes less and less manual.”
New types of chatbots are being built by analysing lots of previous live chat interactions, and some “voice bots” do the same with speech-to-text transcripts.
These new types of chatbots are being built by analysing lots of previous live chat interactions, and some “voice bots” do the same with speech-to-text transcripts.
Just remember, though, to make these work, you’re going to need data. A lot of data!
3. Chatbots Are Best Used in Predefined Scenarios
While it’s chatbots that are currently taking all the headlines in the AI space, as they fit the whole robots taking over human jobs rhetoric, robotic process automation (RPA) is slowly taking over back-office duties in many contact centres.
Why? Because these conscious tasks that humans undertake are easy to encode, such as following processes, pattern recognition and mathematical interpretation. Whereas unconscious tasks, like conversation, are much more difficult.
When it comes to chatbots, those which work the best often handle just one type of query, where rules can be easily encoded.
So, when it comes to chatbots, those which work the best often handle just one type of query, where rules can be easily encoded.
Dr Nicola Millard, Head of Insight & Futures at BT, says: “Machines are very good at doing things where there is process and where there is data, where there are rules. They’re not very good at things that are messy, complex and emotive.”
“That’s why my personal favourite example of a chatbot is Margot (Lidl’s wine bot), because it’s focused on doing one thing really well – food and wine combinations. You can’t ask Margot about anything else.”
So, if you want your chatbot to cover more than just one query type, it might be best to look into implementing multiple bots, each tasked with handling a different contact reason.
4. Multiple Chatbots May Be Necessary – Each With a Different Language Style
It’s important to remember that the mental state of the user is critical to the customer experience. But the mental state of the customer will vary from one contact reason to another.
Sharing an example, Thomas Hebner explains: “We worked with one airline which had a very fun brand and they were building a voice system, but it would only be active when hold times were high and it would only be used for flight status enquiries.”
“Recognising this was a highly stressful situation, they completely went off-brand with their voice – so they picked a very conservative persona. In a situation when customers just wanted a quick answer, efficiency was key.”
So while you may want a bot persona that matches your brand, in some scenarios that just won’t work. Tone needs to be relevant to the query type, but robots cannot alter their tone from one conversation to another, as it’s impossible to encode social skills.
The best chatbots have a predefined purpose and if you want them to handle multiple query types, you may have to employ multiple bots – each with an appropriate way of speaking.
Again, this is why the best chatbots have a predefined purpose and if you want them to handle multiple query types, you may have to employ multiple bots – each with an appropriate way of speaking.
Also, just a quick reminder not to offer a bot to handle every type of contact. It doesn’t work for things like complaints, as chatbots aren’t good on emotion and you have to ask yourself how much a customer would genuinely appreciate an empathetic response from a machine that can’t process feelings.
5. No Chatbot Can Detect Sarcasm
Reports from a number of sources have, over the past couple of years, reported that robots are now able to report sarcasm. So they have a certain amount of emotional intelligence, right? Nope. The fact is that these reports can be very misleading.
Take The Telegraph’s 2017 article “Oh, great. Robots can now tell when people are being sarcastic”, which reports that MIT researchers have created a tool to detect sarcasm in social media posts. It’s only when you get a few paragraphs into the article that you find out the tool only interprets how emojis are used in posts – not the actual words.
Reports like this are why Thomas Hebner feels the need to set the record straight: “Anyone that tells you that they are detecting sarcasm is telling you stories; even certain people don’t understand sarcasm.”
“I went to graduate school and in my classes I was the only native-English speaker, so we developed a rule that when I was using sarcasm, I would have to raise my hand.”
“As humans we don’t even know how to figure that out, so being able to detect sarcasm or even someone’s emotional state are still things that we’re all building towards.”
However, just because chatbots cannot detect sarcasm, it doesn’t mean that they cannot flag up when a customer uses certain emotive words.
If paired with sentiment analysis, it is possible to detect whenever a customer uses an emotive word such as “promise”, to then pass the contact through to a human advisor, who can then make the promise instead of the chatbot. This better reassures the customer.
6. Lots and Lots of Data Is Required
There’s no easy answer to this one. Sorry! But ensuring that you have lots of data from previous contact centre interactions is key.
As Dr Nicola Millard tells us: “When I’m drawn into conversations about bots, firstly I ask: why do you want to use one? So we can define its purpose. And secondly, I ask: what data have you got to fuel it? Because often, enterprise data is really messy, so there can be quite a big problem there.”
In truth, you can employ the best technology available on the market, but if it’s not fed with lots of relevant data, it cannot “learn”, making its key function obsolete.
As John Finch, Product Marketing & Content Executive at RingCentral, says: “The key to a successful chatbot is the data behind it. For the majority of chatbots, the best data will be in the company’s internal data, but there is also external and unstructured data that can really help you to develop the right chatbot for your organisation.”
With there being so much value around sourcing this data, many vendors now partner with Google, as 400 million people are using Google Assistant, which provides vendors like RingCentral with a huge database that supports their chatbots and other AI technologies.
7. Chatbots Are Now Being Used More for “Agent Assist”
For the time being, there has been a notable shift in the way that some companies are approaching chatbots, with a few organisations employing them to work side-by-side with a human advisor as opposed to being a customer-facing “channel”.
A key reason for doing this is to “train” the chatbot to learn from the customer–advisor interactions, while offering the advisor support, by pulling forward useful information from contact centre systems, such as the knowledge base, to help move the conversation along.
So the chatbot is able to listen in on the conversation, possibly across every channel, suggesting answers and offering guidance. But this is just one take on “agent assist”.
The other take is in those circumstances when a call gets escalated from a front-end chatbot to an advisor. How can the bot still help the advisor then?
Meghan Keough, Global VP Corporate at 8×8, suggests that “the chatbot needs to show all the relevant context to the agent who takes over that contact. This includes the conversation transcript and customer’s information.”
“This means that the agent is starting with the best information possible, so the relevant data needed to answer the query is spoon-fed to the agent.”
Chatbots are a marketeer’s dream. The benefits are so easy to picture. But with all of this excitement, it appears we’ve all got a little carried away.
We hear all of the bold new claims about the latest innovations in these bots, when the reality – in some cases – is really quite different:
- “Our bot is virtually plug in and play” – Well then, it’s not going to be very good.
- “Our bot can handle a variety of contact reasons” – … in the same tone of voice that’s appropriate for some interactions and not others.
- “Our bot can detect sarcasm” – Unless your DeLorean took you years into the future and you stole some incredible new technology, it won’t.
However, despite all of these myths that circle the technology, there are many use cases where chatbot implementation has been a success.
For example, Margot – the Lidl wine bot – has had many plaudits. It is arguably the best example of how a chatbot works best in a predefined role and, in the future, it may soon become common practice to give each bot a job description when planning its implementation.
Yet for now, many organisations have chosen to implement chatbots in the form of “agent assist”. This uses AI to search the contact centre knowledge base and other systems to produce proactive knowledge suggestions, while tracking the customer conversation, to help the advisor move the conversation along.
To find out more about AI in the contact centre, read our articles: